An Econometric Model for Measuring System-level impacts of AI on United States Power Grids

Dana Annalise Golden, Aruna Balasubramanian, Niranjan Balasubramanian

Keywords: Electricity, Energy, Artificial Intelligence, Economics, Power, AI Efficiency, LLMs

TL;DR: We utilize econometric techniques to examine the impacts of AI models on demand for power and power quality in the United States.

Abstract: Data centers are a major source of power demand growth in the 21st century. Artificial Intelligence has accelerated this trend with power demand by data centers growing from 1.9\% of total US power demand to 4.4\% of US power in only six years. While anecdotal evidence suggests that AI data centers are using enough power to have substantial impacts on the U.S. power grids, there are no systematic studies to quantify these effects. We utilize econometric techniques to determine the impact of AI model training and inference on consumer electricity quality and fossil fuel power demand. We find significant reductions in power quality and significant increases in power demand near data centers both immediately before and immediately after the publication of AI models. The largest impact worsens power quality equivalent to an additional .5-1 power outages per year. We further show these estimates can also be used for counterfactual analysis to assess impacts of scaling for future model development.

Primary Area: infrastructure, software libraries, hardware, systems, etc.

Submission Number: 13788